Safety4h ago
Safety
Claude Opus 4.6 Faces Elicitation Challenge
A user on the AI Alignment Forum launched a challenge to test Claude Opus 4.6 on Ancient Greek. The experiment targets sycophancy, where the model mirrors user errors rather than providing correct answers. This test highlights the difficulty of unsupervised elicitation. Practitioners can use these results to gauge model honesty in niche academic domains.